# Cross-DB Connection Layer — Structure Sample

This directory shows the **shape** of the cross-DB connection
artifacts that bridge Fragrantica and Parfumo. Files here are
**structural samples** — they demonstrate column layouts, formats,
and content style without trying to be join-able with the 10+10
perfume samples in `../fragrantica/` and `../parfumo/`.

For the full cross-walk dataset (80,968 perfume pairs + 6,522 brand
pairs + complete overlap stats), see
`cross-source-full-2026-05-28.zip` from the $400 / Annual / Lifetime
tier.

---

## Files in this sample

| File | Content | Full release equivalent |
|---|---|---|
| `matched_pairs_sample.csv` | First 30 of 80,968 F.pid ↔ P.pid pairs | `matched_pairs.csv` |
| `brand_matches_sample.csv` | First 30 of 6,522 brand pairs | `brand_matches.csv` |
| `field_equivalence_map.csv` | Full schema-level F↔P column equivalence (206 rows) | same — small enough to ship whole |
| `accord_overlap.json` | 15 matched accord pairs + F/P-only counts | same |
| `notes_overlap.json` | 1,690 direct + 7,977 species-of summary | same |
| `perfumers_overlap.json` | 1,669 perfumer pairs + Jaccard | same |
| `dictionary_overlap_stats.json` | Aggregate stats across notes/accords/perfumers/brands | same |
| `overlap_stats.json` | Top-level Jaccard summary | same |

---

## File schemas

### `matched_pairs_sample.csv` — F ↔ P perfume cross-walk

Columns: `frag_pid | parf_pid | match_pass | brand_relation | confidence`

```
frag_pid|parf_pid|match_pass|brand_relation|confidence
30|728|exact|same-brand|exact
762|727|exact|same-brand|exact
```

- `match_pass`: how this pair was matched (`exact`, `brand_fuzzy`, `token-subset`, etc.)
- `brand_relation`: `same-brand` / `cross-brand` / `unknown`
- `confidence`: `exact` / `confident` / `low` — based on G1 stratified audit
- Method: Phase 2B Fellegi-Sunter probabilistic record linkage
- Precision in full release: ~99% (G1 audit)

### `brand_matches_sample.csv` — F ↔ P brand cross-walk

Same shape as matched_pairs but at the brand level.

### `field_equivalence_map.csv` — schema-level column equivalence

Columns: `f_field | f_file | p_field | p_file | relation | note`

`relation` values:
- `equivalent_same_format` — drop-in interchangeable (21 rows)
- `equivalent_diff_format` — same concept, different encoding (31 rows)
- `partial_overlap` — overlaps for some uses, diverges for others (7 rows)
- `f_only` — F has it, P doesn't (59 rows)
- `p_only` — P has it, F doesn't (88 rows)

Used to know "does F column X have a P counterpart, and if so what's
the encoding relation?" — see also `field_equivalence_map.md` in the
full release for the long-form reference.

### `*_overlap.json` files

Each describes one entity type's cross-DB overlap:
- counts (matched, F-only, P-only)
- Jaccard index
- method (Phase 2B / Phase 2.5 / species-of taxonomy)
- snapshot caveats (e.g. perfumers_overlap was computed on F=2,893
  snapshot before v5.4 added 75 perfumers)

---

## How to use this for a join (example)

```python
import pandas as pd

# Load the cross-walk
mp = pd.read_csv("cross/matched_pairs_sample.csv", sep="|")
print(f"Pairs in sample: {len(mp)}")  # 30

# Filter by confidence
exact = mp[mp.confidence == "exact"]
brand_fuzzy = mp[mp.match_pass == "brand_fuzzy"]

# In the FULL release, this scales to:
#   80,968 total pairs · ~99% precision (G1 audit)
#   Use frag_pid / parf_pid to join the per-DB CSVs
```

---

## What this sample does NOT include

- Full `matched_pairs.csv` (80,968 rows / 3.2 MB) — in full release
- Full `brand_matches.csv` (6,522 rows / 323 KB) — in full release
- `field_equivalence_map.md` (27 KB, human-readable long-form) — in full release
- Per-row audit decisions or scoring features — in full release

Cross-walk methodology details are in `cross/README.md` of the full
bundle.

---

## Notes on the F + P sample picks

The 10 F perfumes and 10 P perfumes in `../fragrantica/` and
`../parfumo/` were chosen independently — each side picked the most
populated rows on its DB, top by rating votes. They don't necessarily
join through `matched_pairs.csv` (the cross-walk shown above is a
structural sample, not curated to the 10+10).

For a cross-walk demo that links F and P at the perfume level, see
the full release.